Claude Fable 5 Is Back: What It Actually Costs to Run

Claude Fable 5 is back, and it did not just quietly reappear on the API. It was pulled entirely on 12 June 2026, following a US government export control order, and restored on 1 July 2026 across the whole Claude lineup: the API, Claude.ai, Claude Code, and Claude Cowork. For businesses building on it, or paying for it through a subscription, the question that actually matters now is not whether it is capable. It clearly is. The question is what it costs to run at real volume, whether you are paying by the token or by usage credits, and whether that cost is justified for the task at hand. Here is the full breakdown.

Why It Was Pulled in the First Place

The suspension was not a rollout hiccup or a capacity problem. Amazon researchers had found a way to prompt around Fable 5's safeguards to get the model to identify software vulnerabilities, and in at least one case the model produced code demonstrating how the vulnerability it found could be exploited. That report reached the US government, which responded on 12 June 2026 by applying export controls to both Claude Fable 5 and Claude Mythos 5, restricting access to foreign nationals. Anthropic had no reliable way to verify a user's nationality in real time, so rather than risk a breach of the order, it suspended both models entirely, for every user, everywhere.

Anthropic used the suspension to train an improved safety classifier that specifically targets that bypass technique. By its own account, the new classifier blocks the exact method described in the Amazon report in more than 99 per cent of cases. That classifier did not disappear once access was restored. It is now a permanent part of how Fable 5 behaves, and it is the reason the model can return a refusal on a request rather than an answer, a behaviour covered in the billing section below. The export controls were lifted on 30 June 2026, and Fable 5 came back online for everyone the next day.

What Claude Fable 5 Actually Is

Claude Fable 5 is built for the most demanding reasoning and the longest horizon agentic work: multi step research, autonomous coding runs, and tasks that require sustained context over many tool calls. It ships with a 1 million token context window as the default, and can return up to 128,000 tokens of output in a single request.

One detail that matters for cost planning: thinking is always on. There is no way to switch it off, and the depth of that thinking is controlled with an effort setting rather than a fixed token budget. Because reasoning happens whether or not it is shown to you, it is billed as output tokens either way. That is worth remembering before you compare a visible response length to the bill you actually receive.

Where It Is Available, and What That Means for the Bill

Fable 5 is not an API only model, and it is worth being precise about that since it changes what "cost" even means depending on how a business reaches it. Since 1 July 2026 it has been available globally on the Claude Platform (the API and its underlying infrastructure), Claude.ai, Claude Code, and Claude Cowork. Cloud deployments through Amazon Bedrock, Google Cloud, and Microsoft Foundry are being re-enabled separately, so check current status there before assuming it is live on every platform your business uses.

That distinction matters most for the billing model. Businesses calling Fable 5 through the API pay the per token rates in this article directly. Businesses on a Claude subscription, Pro, Max, Team, or select Enterprise plans, had Fable 5 included for up to 50 per cent of their weekly usage limits through 7 July 2026. After that, further use on those plans draws on usage credits, which is Anthropic's metered, consumption based billing layered on top of the subscription. In practice, that means subscription users land on the same underlying logic as API pricing once the included allowance runs out: heavier Fable 5 use is charged by what is actually consumed, not folded quietly into a flat monthly fee. The token by token numbers below apply directly to API use, and are the same economics that determine what a usage credit actually costs once a subscription's included allowance is used up.

The Headline Numbers

How It Compares to the Rest of the Claude Lineup

Fable 5 sits at the top of the price range, and it is worth seeing that against the other current models before deciding where it belongs in your workflow.

Model	Input, per million tokens	Output, per million tokens	Context window
Claude Fable 5	$10.00	$50.00	1 million tokens
Claude Opus 4.8	$5.00	$25.00	1 million tokens
Claude Sonnet 5	$3.00 Introductory $2.00 through 31 August 2026	$15.00 Introductory $10.00 through 31 August 2026	1 million tokens
Claude Haiku 4.5	$1.00	$5.00	200,000 tokens

Fable 5 costs twice as much as Opus 4.8 on both input and output, and more than three times the standard rate for Sonnet 5. For a business processing large volumes of routine requests, that difference compounds quickly. It only makes sense once the task genuinely needs the extra capability.

Cached Tokens Change the Maths Considerably

Most real workloads reuse the same content across many requests: a system prompt, a long reference document, a set of tool definitions. Prompt caching lets you pay a reduced rate for that repeated content instead of the full input price every time.

Cache behaviour	Price, per million tokens	Notes
Cache read	$1.00	About a tenth of the standard input rate
Cache write, 5 minute TTL	$12.50	1.25 times the standard input rate
Cache write, 1 hour TTL	$20.00	Twice the standard input rate

Consider a 50,000 token contract that gets analysed once and then asked ten follow up questions. Without caching, every single question pays the full rate to re read that document: roughly $0.50 each time, or $5.00 across ten questions just to keep re reading the same text. With caching in place, the first read is billed as normal (or slightly more if it is written to the cache), and every question after that reads the cached copy at roughly a tenth of the price, about $0.05 each. Nine of those ten reads drop from $0.50 to $0.05, cutting the document reading cost by around 90 per cent.

The token price gets the headlines. The caching strategy is what actually decides the bill.

A Worked Example: One Quick Task vs a Long Agentic Run

To put the numbers in context, here are two realistic scenarios.

A single short task

Drafting a customer email from a brief brief: say 1,000 input tokens and 1,000 output tokens. That works out to $0.01 for the input and $0.05 for the output, so roughly $0.06 per email. Run a few hundred of those a month and the total stays modest. Run tens of thousands, and the model choice starts to matter a great deal.

A long autonomous session

Fable 5 is built for turns that can run for many minutes on hard problems, gathering context, building, and checking its own work before finishing. Picture a research or coding session that consumes 300,000 input tokens across the run, much of it repeated context, plus 40,000 output tokens including the hidden reasoning. Without any caching, that is $3.00 for input and $2.00 for output, about $5.00 for the session. With the repeated context served from cache instead of paid at full price, the input side can realistically drop to well under a dollar, bringing the whole run closer to $2.50 to $3.50. The difference between a well cached session and an uncached one, at this scale, is not trivial.

Practical Levers for Managing the Bill

• Set the effort level deliberately. Thinking is always on for Fable 5, and it is billed as output. Use a lower effort setting for routine steps and reserve the higher settings for genuinely hard reasoning.
• Structure prompts for caching. Put stable content, like system instructions, reference documents, and tool definitions, ahead of anything that changes per request, so it can be reused from cache rather than repriced every time.
• Use batch processing for anything that is not time sensitive. Non interactive workloads processed asynchronously are billed at half the standard token price.
• Do not default to the most expensive model. Routine drafting, classification, and lookups often perform just as well on Sonnet 5 or Haiku 4.5, at a fraction of the cost. Save Fable 5 for the work that actually needs it.
• Check your data retention settings first. Fable 5 requires 30 day data retention and is not available under a zero data retention configuration, so this is worth confirming with your compliance requirements before committing budget to it.
• Refused requests can cost nothing. If a safety check declines a request before it produces any output, that request is not billed at all. Only a decline partway through billing for the output already produced up to that point.
• If you are on a subscription rather than the API, watch your weekly allowance. Pro, Max, Team, and select Enterprise plans included Fable 5 usage for free up to a limit through 7 July 2026. Beyond that, heavier use is metered through usage credits, so a team that leans on Fable 5 daily should track consumption the same way an API customer tracks tokens.

Should Your Business Use It

Fable 5 earns its price on long horizon agentic work, complex multi step reasoning, first pass builds of well specified systems, and high stakes document or code review, the kind of task where getting it right the first time is worth more than the extra spend. For everyday content generation, simple customer replies, or high volume low complexity work, a cheaper model in the Claude lineup will usually deliver a comparable practical result for a fraction of the token cost.

The right approach for most businesses is not picking one model for everything. It is matching the model to the task, and having the caching and effort settings tuned so you are not paying flagship prices for routine work.